Research On Deep Learning Models Of Pedestrian Detection For Road Scenes

In recent years,with the rapid development of science and technology,autonomous driving,unmanned cars and other technologies have become highly popular,which has been the focus of attention of all walks of life.At the fourth session of the 13 th National People’s Congress,which just ended this year,many proposals on "Autonomous Driving","Unmanned Vehicle" and "Intelligent Vehicle" have sprung up.However,the above technology is not mature,so terrible accidents often occur.The pedestrian detection technology is not accurate and not timely which leads to such a result.Therefore,improving pedestrian detection technology under road scenes,increasing detection accuracy and reducing detection miss rate have become an urgent problem to be solved in autonomous driving technology.In view of the challenges faced by pedestrian detection tasks,this thesis puts forward the following research points:1.This thesis analyzes the existing problems of small-scale pedestrian detection in pedestrian detection.Considering the differences of network learning ability for pedestrians of various scales,it will study from the multi-scale aspect.Based on the two-stage Faster R-CNN detection deep learning method,multi-scale feature fusion strategy and multi-scale receptive field RPN strategy are proposed.In this thesis,a multi-scale feature fusion model is designed to improve the learning ability of the model for low-level features;Secondly,a multi-scale receptive field(RPN)model is proposed to explore the potential of RPN model;The experimental results show that the design of the model can improve the performance of pedestrian detection task and the effectiveness of small-scale pedestrian detection.2.Considering that the small-scale pedestrian features in pedestrian detection are less concerned by the network,and the corresponding features get less attention from the network,which leads to less learning features.Therefore,from the perspective of attention mechanism,based on the one-stage detection algorithm of Retina Net,an Adapted Retinanet model for multi-scale pedestrian detection task is designed.Based on the one-stage pedestrian detection model,this thesis constructs three improved strategies.Firstly,the role of CBAM attention mechanism in small-scale pedestrian detection is fully explored;Secondly,the BN normalization method in Retina Net is transformed into GN;Finally,DIo U Loss is used to replace the loss function.The experimental results show that the above design can improve the detection performance.3.Based on the above work,this thesis further discusses the impact of data augmentation strategy on pedestrian detection performance.Specifically,the advantages and disadvantages of Oversampling method and Copy-Pasting method in pedestrian detection model are compared.Through the experimental analysis,the applicability of the method in this chapter to solve the multi-scale pedestrian problem is evaluated,and the effectiveness of the method in this chapter for small-scale pedestrian detection is verified.